JPMorgan AI Models Hit Sub-250 Microsecond Speeds
JPMorgan Chase's trading floor AI models now process 3 billion market events daily, with individual inference tasks completed in under 250 microseconds. The system is required to maintain full audit trails, setting a new industry benchmark for combining ultra-low-latency performance with stringent compliance in financial AI infrastructure.
- JPMorgan's AI-powered trade execution platform, known as LOXM, has been in development since at least 2017 and was trained on billions of historical transactions to optimize execution speed and pricing. - The firm's broader AI infrastructure is managed on a proprietary, cloud-flexible platform called OmniAI, which is used by data scientists and engineers across all lines of business to accelerate the deployment of AI/ML applications. - In high-frequency trading, the time difference between a successful trade and a failed attempt can be as narrow as 5-10 microseconds, highlighting the competitive necessity of the sub-250 microsecond speeds JPMorgan has achieved. - To manage regulatory requirements, the bank's compliance systems use AI to reduce false positives in transaction monitoring by 30-45% and cut operational spending on KYC and AML workflows by up to 25%. - The daily processing load for the bank's systems exceeds one exabyte of data, which includes structured data, unstructured data, and voice and video files. - In a recent internal reorganization to accelerate AI adoption, JPMorgan named Guy Halamish as COO of its commercial and investment bank, giving him direct oversight of data and AI strategy across the division. - The underlying infrastructure for these AI systems is a hybrid model that includes private data centers, colocation facilities, and public cloud deployments, utilizing technologies like Kafka for data streaming and gRPC for low-latency model serving. - Beyond trading, the bank has more than 450 generative AI use cases in motion, including an internal coding assistant for developers and LLMs for summarizing research and client portfolios.
